AI-Driven Cooling Load Forecasts in Dense Urban Environments

Abstract Electrification of space heating and cooling, together with more frequent and intense temperature extremes, is shifting when and where peak electricity demand occurs, increasing stress on power systems during weather-driven load surges. At the same time, utilities increasingly need neighborhood- and substation-scale load forecast, because peak demand in dense cities is strongly modulated by urban meteorology and building–atmosphere interactions that are not resolved by conventional, territory-averaged forecasting workflows. Here, we present an AI load forecasting model for dense urban environments, focusing on the New York City metropolitan region, that predicts spatially distributed cooling load directly from meteorological fields. The model is trained using high-resolution (1.3km) uWRF simulations that couple an urban canopy parameterization with a building energy model, enabling physically consistent HVAC load targets across the urban landscape. Using a convolutional encoder–decoder (U-Net) architecture, we map hourly near-surface temperature, relative humidity, and heat index to hourly spatial cooling-load fields and evaluate performance across multiple summers. The model reproduces the dominant diurnal and synoptic variability in urban cooling demand and shows strong agreement when aggregated to ISO zone scales, while skill decreases during the most extreme peak hours, consistent with the amplified sensitivity of cooling load to high-temperature conditions. These results demonstrate a practical pathway for learning spatially explicit urban load behavior from physically based simulations, providing an operationally efficient complement to computationally intensive urban climate simulations and building-energy modeling.